Join us on March 2nd for an ActiGraph webinar:
Wearable Data Gone Awry: Cautionary Tales from the Clinical Research Trenches.Register Now
Physical Activity Change in an RCT: Comparison of Measurement Methods
- Published on May 7, 2019
We aimed to quantify the agreement between self-report, standard cut-point accelerometer, and machine learning accelerometer estimates of physical activity (PA), and examine how agreement changes over time among older adults in an intervention setting.
Data were from a randomized weight loss trial that encouraged increased PA among 333 postmenopausal breast cancer survivors. PA was estimated using accelerometry and a validated questionnaire at baseline and 6-months. Accelerometer data were processed using standard cut-points and a validated machine learning algorithm. Agreement of PA at each time-point and change was assessed using mixed effects regression models and concordance correlation.
At baseline, self-report and machine learning provided similar PA estimates (mean difference = 11.5 min/day) unlike self-report and standard cut-points (mean difference = 36.3 min/ day). Cut-point and machine learning methods assessed PA change over time more similarly than other comparisons. Specifically, the mean difference of PA change for the cut-point versus machine learning methods was 5.1 min/day for intervention group and 2.9 in controls, whereas it was ⩾ 24.7 min/day for other comparisons.
Intervention researchers are facing the issue of self-report measures introducing bias and accelerometer cut-points being insensitive. Machine learning approaches may bridge this gap.
- Nelson, Sandahl H.
- Natarajan, Loki
- Patterson, Ruth E.
- Hartman, Sheri J.
- Thompson, Caroline A.
- Godbole, Suneeta V.
- Johnson, Eileen
- Marinac, Catherine R.
- Kerr, Jacqueline
American Journal of Health Behavior